using physical threads

[DiamonDinoia]

Attempting to fix #596

Two main reasons:

• Vector unit are per physical core.
• Manually vectorized code is (also the FFT) are memory bound.

Using HT cores impairs performance due to extra scheduling and puts unnecessary strain onto the memory controller

[mreineck]

Just a small word of caution: this will always report the full number of physical cores on the system, independent of possible restrictions on the executable itself. On computing center nodes, the system may limit a job to a part of the available hardware (e.g. via taskset); if this happens, the code here will overestimate the number of usable cores.
On systems with power and efficiency cores the code will probably (this is just speculation, I haven't tested it) return the sum of both core counts, which may also be larger than the amount of cores that should actually be used.

I'm not saying that the code should be changed (I wouldn't know how...), but this could be worth remembering in case there are reports about weird behaviour in the future.

[DiamonDinoia]

Yeah, this emerged from some discussions with @blackwer yesterday. I'll see if I find a way to figure out the active cores.

[blackwer]

I was going to suggest looking at hwloc for inspiration, and it looks like you've found a very similar solution that they did. For posterity:
https://github.com/open-mpi/hwloc/blob/1779eae26f2f55a510e9d73cb648b68e71d9e6d4/hwloc/topology-linux.c#L1036

[mreineck]

I was going to suggest libpthreads ... it seems this functionality is available everywhere these days :-)

[DiamonDinoia]

Where is the functionality in pthread? I didn't go the route of hwloc because I read somewhere that is an issue on maybe freebsd? This seemed more portable to me

[mreineck]

In ducc I'm using

size_t available_hardware_threads()
  {
  static const size_t available_hardware_threads_ = []()
    {
#if __has_include(<pthread.h>) && defined(__linux__) && defined(_GNU_SOURCE)
    cpu_set_t cpuset;
    CPU_ZERO(&cpuset);
    pthread_getaffinity_np(pthread_self(), sizeof(cpuset), &cpuset);
    size_t res=0;
    for (size_t i=0; i<CPU_SETSIZE; ++i)
      if (CPU_ISSET(i, &cpuset)) ++res;
#else
    size_t res = std::max<size_t>(1, std::thread::hardware_concurrency());
#endif
    return res;
    }();
  return available_hardware_threads_;
  }

which is very close to your getAllowedCoreCount function, but it gets the affinity masks from libpthreads instead of the GNU libc. It might be more portable, but I'm not sure.

[ahbarnett]

Seek opportunities to delete code :) Thanks